Designing for AI-as-a-tool

"AI is a tool to increase our capabilities, not to replace us."

―Yann Le Cun

Artificial intelligence is everywhere nowadays. It is a shock to the system in not only the tech world, but also society in general. Like a stone tossed into a pond, causing ripples and waves across the surface, AI has sparked many different reactions. Stories about how employees of a tech giant have been wasting tokens in order to meet their new AI use quota. Calls on social media to halt data center construction for environmental concerns. Fierce debate over whether or not AI-generated art is art. Mass layoffs from corporations who dreamt of replacing their workforce with AI agents, and the disastrous consequences. AI agents identifying tissue features that may predict future breast cancers. AI agents losing someone's lifesavings trading stocks.

Today, the debate has grown to become much more nuanced than being pro or anti AI use. Today, it is paramount for us to figure out how to coexist with AI, how to use it to our advantage, to enhance our productivity, and more importantly, our quality of life. How do we make AI work for us, and vice versa?

"AI is a powerful tool, but it's just a tool. It's up to us to decide how to use it."

―Yoshua Bengio

As a UI designer, I am interested in exploring design practices to help our users better command their AI tools to achieve their goal.

There is this famous case study for a design practice we designers call "human-in-the-loop design". During WW2, Bell Labs developed the M9 gun director, a machine that continuously calculated trigonometric firing solutions for anti-aircraft weapons. It was paired with the SCR-584 radar and 90mm anti-aircraft guns. The system was an advanced piece of combat engineering, but it refused to function without 2 human operators: a radar operator, who has to acquire and confirm the target, and a gunner, who has to physically pull the trigger to fire upon an aircraft. This was a deliberate choice. The military wasn't going to hand a machine the trust to decide which crafts to shoot down and the authority to kill human combatants.

Machines and humans make a good team. They leverage each other's strengths and complement each other's drawbacks. With human-in-the-loop design, designers place humans as an important component in a system that can do amazing things.

Upon deliberation, I've come up with these 4 points to help guide the design of AI tools which users can wield with efficiency, clarity, and accountability.

Predictability

Any interface which we use to interact with AI should always remain predictable. That is to say, AI should not alter the workspace.

A negative example of this would be the "smart tabs" that try to predict what the user would likely want to do next based on habit, time, location, and other data. 

Take, for example, the home screen of a smartphone that features a smart app drawer. The user uses the music app daily. So much so, that the user has developed muscle memory to find and open the app, navigate to their "liked songs" playlist, and hit play; all without any conscious thought from the user. But now, since the smart app drawer has been implemented, the music app is moved around in the app drawer each time the user wants to find it. The user has to re-find their music app each time they want to listen to music. This increases the mental load of the user and interrupts their flow. Even if the music app was left in the same spot sometimes, the user would still have to pause to think about it.

For a human, cognition is embodied and in the world. Creative minds have messy desks, and chefs swear by their mise-en-place. We offset our thinking into our workspace. In a user interface, once the user has become accustomed to the interface, that interface becomes an extension of the user, not the machine. To let the AI alter the workspace is to encroach on the user's natural flow, to step on the user's toes.

Transparency

To most people, AI is a blackbox. We have no idea what's really going on under the hood. This is not a big deal, as the interface is here to help bridge the gap. To ensure that the user understands the tool they are using, the interface must communicate to the user what the AI is doing.

This is something online AI tools already do. Things like "searching web…", "running code…", and "reading file…" keep the user in the loop while their response is being generated. This is a good start. I believe we can bring it a step further by communicating abstract thought actions to the user as well.

Artificial intelligence is often referred to as "hallucination engines". It does not think logically like many people mistake it to do. How do we show the user the actual process behind the curtains?

In the wild west days of interactive design, computer scientists leaned on metaphors. They used metaphors to link computer functions to physical things in the real world. You have "files" on your "desktop", which you can use "tools" in your "tool bar" to make changes to. We can do something similar with abstract thought actions. Words like "associating", "retrieving", "imagining", can communicate to the user the actions being taken by the AI.

Pulling the trigger

In the previously mentioned example of the M9 gun director computer, a gunner is responsible for pulling the trigger. An accountable human being has to approve the actions of the machine, and commit to the action that materializes the work of the computer into the real world.

Any work an AI does should be kept in a quarantined environment, until a responsible human chooses to commit the changes. Upon this commitment, the human operator should assume the responsibility of the consequences of the action.

Version control

If you only make small tweaks to alter an AI's result with word prompts, it is not rare to be thrown in a loop wasting tokens. Prompting an AI to give you the results you want, like most work we do in the tech space, is iterative. You add or subtract certain parameters to see if it will yield a better result. Thus, like any system for managing iterative work, a system of version control should be put in place. There should exist a way to return to previous results with different parameters.

Of course, you can simply use git with your AI agent. But this is complicated and not accessible to the general public. We can add a graphical interface to show the different branches of AI generated results.

Conclusion

Ultimately, designing for AI as a tool means designing for human agency. Predictability preserves the user's flow, transparency preserves understanding, human approval preserves accountability, and version control preserves exploration. None of these principles make AI less powerful; they make it more usable, trustworthy, and aligned with the people it is meant to serve. The goal is not to build systems that act independently of us, but systems that amplify what humans already do well while compensating for our limitations. The most successful AI products will be those that treat intelligence as a partnership rather than a replacement. The future of AI should not be one where humans adapt themselves to machines, but one where machines are carefully designed to extend human capability.